Modeling Approximate Computing Techniques

The efficiency–accuracy trade-off of approximate computing spans a diverse array of techniques at both the hardware and software levels. While this diversity is key to the success of approximation research, it also entails considerable complexity in developing and validating new approximation techniques. Researchers must relate their ideas to a vast body of work on approximate computing and invest significant software and hardware implementation effort to validate their hypotheses. These problems slow the pace of innovation and, when they prove too much to bear, risk misdirecting research away from promising ideas. To better understand these trade-offs, we present REACT: a new tool for modeling the efficiency–accuracy space afforded by approximate computing. We first motivate the need for such a tool by discussing the complex space of current approximate computing research. Then, we present REACT, a modeling framework that lets researches rapidly evaluate approximate-computing techniques and lets practitioners understand the potential impact of approximation in their code. REACT combines an application profiler with an energy model and error injection framework to measure the effects of approximation on user-provided applications. We show how REACT can easily model approximation techniques from diverse regions of our classification. To demonstrate REACT’s utility, we use it to evaluate several benchmarks for their amenability to approximation and to individual techniques. Our exploration shows that the energy savings and error vary widely between techniques and applications, and that coarse-grain techniques dominate fine-grain ones.